Unsupervised hyperspectral noise estimation and restoration via interband-invariant representation learning

Abstract

Hyperspectral images (HSIs) acquired from different imaging platforms are inevitably contaminated by multiple types of noise. However, the existing supervised learning based denoising methods often show poor generalizability on data with complex degradation, due to the discrepancy between synthetic training data and real data. Although some unsupervised denoisers have been developed to learn priors on real data, the noise assumptions or image priors in these methods limit their performances. In this paper, an unsupervised noise estimation and restoration (UNER) framework is proposed based on disentangled representation learning, to create an interband representation space that is resistant to noise within a single HSI, i.e., an interband-invariant representation space. Real HSIs are categorized internally into noisy and clean bands by noise intensity estimation. Intraband and interband disentangled reconstruction are designed to train two encoder-decoder modules to learn the interband-invariant representation. Noise patterns are separated from HSIs by transforming noisy bands into the representation space without introducing hand-crafted priors. The estimated noise and clean bands are then combined to train the self-supervised interband information restoration module, thereby exploiting spectral correlation and restoring the latent noise-free data with high information fidelity. In this way, the UNER framework can learn specific priors from real HSIs without clean data and be applied to various scenarios with complicated noise patterns. Noise removal experiments conducted on three airborne hyperspectral datasets representing urban, agricultural, and forestry areas respectively demonstrate the superiority of the proposed UNER framework over the state-of-the-art hyperspectral denoising methods.

Keywords

• Hyperspectral images
• Hybrid noise removal
• Unsupervised learning
• Disentangled representation learning
• Convolutional neural network